smritijha19/Nivi_combined_clarification_set_1

Nivi Clarification Policy Combined v1

Synthetic counterfactual dataset for training a maternal-health clarification policy model.

Task

Given a maternal-health user query and available context, choose exactly one action:

• —ANSWER_DIRECT
• —ASK_CLARIFICATION
• —ESCALATE_DIRECTLY

The model output is a short reasoning trace in <think>...</think> followed by a JSON decision.

Language

User queries may be in English, Hindi, Assamese, or code-mixed language. The reasoning trace and JSON output are in English.

Included families

This combined v1 dataset includes:

• —bleeding in pregnancy
• —preeclampsia-like headache/swelling symptoms
• —abdominal pain in pregnancy
• —postpartum discharge/infection concern
• —hard negatives / non-triage informational questions

Important caveat

This is a synthetic dataset generated from provisional clarification rubrics and scenario cards. It is not final expert-labeled clinical data. Real Nivi-style examples were used only as style and topic anchors, not as public gold-labeled raw data.

Intended use

Initial SFT baseline for a small LLM clarification policy model.